Resources » Frequently Asked Questions

In This Section:

On This Page:


How does Neurala secure images and videos?

  • Images are stored with Amazon S3 security (HTTPS Protocols) with timed expiration URLs

  • By default, data is currently not encrypted at rest.

What browsers and devices does Brain Builder support?

  • Chrome is the only supported browser. Other browsers are not tested widely and could be buggy.

  • Brain Builder was developed for use on Desktop computers. Tablet use is possible but not widely tested.

Can Brain Builder be installed on-premise?

  • Brain Builder is currently a cloud-only solution. On-premise installation is a consideration for future development.

How does AI Video Annotation work?

  • Neurala uses a proprietary learning technique that learns from the first frame to annotate subsequent frames.

How is Brain Builder updated? Is that part of the contract?

  • Brain Builder is a SaaS platform that will get periodic updates. All updates to Brain Builder are included in the contract, which is a benefit of a SaaS product—the product is always getting better! Updates will be communicated to all users and should have no negative impact on users.

How does a customer add a new class to an existing brain or improve its performance with more data?

  • One of the many benefits of our L-DNN technology is that it supports this use model quite well. Users just need to log into the dataset that contains the brain and add the new class and supporting images. While training the new class, Published brains will have no impact. Once the user is satisfied with the performance of the current brain, it can get promoted to staging or Production for use in the field. The training process is almost immediate so adding classes or improving performance is easily done.

How do we receive predictions and integrate into our solution?

  • You can access all your custom brains via an HTTP API. You can also download an iOS,  Android or Linux SDK, and integrate that to your application via an Objective-C, Java, or C++API, respectively. The downloadable SDK has the ability to download an updated brain from Brain Builder, and also to add knowledge (new examples or classes) to the brain locally.

Data In & Data Out

What types of Images and Videos does Brain Builder support?

What formats of data exports are supported?

  • Annotations can be exported in formats for TensorFlow and Caffe.

  • Brain Builder also supports exports in JSON, which can be easily converted to other formats and configurations.

What about other exports, like LMDB, HDF5, TFRecords, YOLO, etc.

  • Brain Builder does not currently support these export formats. After consulting with researchers, there are enough variables in how this is set up that it makes sense for the Brain Builder user to take care of this themselves.

  • The Brain Builder JSON export can be easily converted to other formats with some light scripting work.

Can Brain Builder export videos?

  • Brain Builder exports annotated video frames as individual images. We’ve seen very few clients use videos to train networks, as there are even more video formats, codecs, etc to support than image types. Supporting video formats, therefore, adds a layer of complexity to the training process, or failure point. Training on videos is not standard.

How does the data export process work?

Can you send me a copy of all of my uploaded data?

  • If you require export functionality other than what Brain Builder currently offers, please contact your Customer Success Manager.

Why would you want a color-coded mask vs. a mask for each class per image (black & white)

  • Let’s say you have a dataset of 1000 images, and you have tagged 7 different classes. AI engineers sometimes want to be able to train on just one or two classes and then discard the rest. The latter would give you that option.

What is the difference between Training, Validation, and Test data?

  • These datasets are data provided for each stage of the AI process:

    • Training is used to train the model, i.e. fitting the parameters (e.g., neural network weights and biases, etc.)

    • Validation is used to tune the hyper-parameters (e.g., neural network architectures, kernel functions in SVMs).

    • The test is used to assess the performance (i.e. FPS and accuracy [mIoU and mAP]). The test data set should not be used in the model building process.

Once I’ve downloaded the dataset, are formats 100% ready to be uploaded into Tensor Flow and CAFFE? What’s the next step?

  • Yes. When a researcher downloads the formatted files they would either place them on a server or their local hard drive. Then when they start a training session in TF or Caffe, they would provide the path to the file they downloaded.

Does the system cope well with unbalanced classes?

  • This really depends on the dataset, but in general, balanced is typically better.


What are the deployment options?

  • A brain can be deployed in the cloud or on the edge with our SDK. Edge deployments can be on iOS or Android devices, as well as certain Linux environments.

How fast will Cloud inference be?

  • On average, inference should take less than 1 second. Depending on the tier contracted, this could be as fast as 100 ms.

Can this run without connection to the internet?

  • If the brain is deployed on the Edge, then yes.

What are the minimum hardware requirements for a recognition use case?

  • iOS: iOS 11.0 or later

  • Android: 8.1 (API level 27) or higher (where Android NNAPI is available)

  • Linux: Ubuntu 16.04, 4+ Core 1+GHz ARM or Intel 64-bit CPU, 200+MB RAM, (CUDA 9.0+)

What are the different frame rates for sample hardware in recognition use cases?

  • Learning and inference of a single image take roughly the same amount of time. Some examples:

    • Intel CPU: 12fps

    • nVIDIA 1060: 212fps

    • nVidia TX2 (GPU): 39fps

    • Samsung Galaxy S8: 9fps

Does an API call that bundles several images into a single call consume a single API Inference or are the Inferences counted per image?

  • Each image constitutes an Inference Call and counts towards your overall account limit.

Why can't I promote my Brain?

  • If the Promote button is disabled, it means no additional training has been done since the last time you deployed, and therefore, it is the same Brain. If you wish to retrain your Brain, click Optimize or Retrain button to potentially adjust your Brain score.

Project Management

What are the features and functionality available to different user roles?


What is the support model?

  • Depending on what tier is contracted, the support models may change.

Does Neurala offer consulting services?

  • Yes. Please contact sales to learn more.

Technical (Nerds Only)

Feature Extraction

What features does L-DNN Recognition work well for (i.e. how do I understand what use case is a good one for L-DNN)?

  • The more distinct the features of different classes are, the better L-DNN (or DNN) will perform in prediction accuracy. Conversely, the more similar the inputs across classes, the higher the error rate.


What is the base size of L-DNN Recognition, especially as I consider running it on the Edge? Will I be able to see how the size is affected as I add new training data? How much memory do I need to keep free on my processor?

  • The total size requirement of the SDK code and the L-DNN brain is around ~25MB on disk. Adding knowledge grows this very slowly, by kilobytes at a time. The memory growth slows as more examples are learned. About 200MB of free RAM should be plenty at run time.

Other DNN Roadmap

What other neural network architectures will Brain Builder offer?

  • We have an exciting roadmap ahead of us and we are always evaluating what the best approaches to solving customer problems are. If you think that our product is not currently solving your problems, feel free to send us feedback at

Image Input/Output

After I upload my imagery into Brain Builder, how is it modified? How is it modified for processing/inference by L-DNN? When I run inference on the L-DNN are there size inputs I can adjust for? What is the maximum? What is the minimum? How does that impact processing time?

  • Image copies are optimized for L-DNN use automatically in Brain Builder as a post-upload task. No user actions are required.

Security Measures


  • At rest, access is limited to specific servers through IAM roles. Engineers can access assets at rest with secret keys.

  • In transit, all transfers use SSL (HTTPS).

  • The database gets backed up every day.

Disaster Recovery

  • Our EC2 instances and most infrastructure runs in multiple AZs.

Employee Security Policy

  • Only employees with Administrator accounts on AWS can access customer assets.

  • Only needed employees can access our production database.

  • All employee keys and accounts are revoked at the time of termination.

Have a question you don't see here? Email and our team will be happy to answer!

Next: Glossary >>